I have been asked a couple of times for a script that would reproduce the results in the time series forecasting series. I finally managed to do it. In the process I found out that a couple of the queries needed to be tuned:
- In the airline example described in Part 2, the normalization shift and scale parameters were computed using the whole data. A better methodology would be to use only the training data for computing these parameters. This alternative normalization scheme also impacts the results for the airline example described in Part 3 as the latter builds on the transformations described on Part 2
- The results for the electric load example in Part 3 did not reflect the forecasting query. They reflected an earlier version of the query. The correct results are a little better
I collected the different steps for both examples in Part 3 into a single script: code_ts_script.sql. The data for the airline and electric load examples can be downloaded from here: data_airline.cvs and data_electric_load.cvs.
I have also made some edits to the Part 2 and Part 3 posts to reflect the points described above.